This research project – a collaboration between data scientists and climate scientists – will develop machine learning solutions to the analysis of the exploding amounts of climate data in order to deliver breakthroughs in climate research, whilst training the next generation of scientists.
Climate change represents one of the most pressing challenges afflicting our societies, with governments worldwide, non-governmental organizations, and other stakeholders placing increasing emphasis on measures to successfully tackle and mitigate its adverse effects. A pressing scientific hallenge relates to the need to understand aerosol-cloud interactions because these currently represent one of the major sources of uncertainty in determining the radiative forcing responsible for climate change projections (Intergovernmental Panel on Climate Change 2021 - IPCC ). One way of reducing the uncertainty involved in determining the radiative forcing of climate change is by understanding the interaction between aerosols, clouds, and precipitation processes. Understanding the complex interactions within the Earth (i.e., the interaction between aerosols, clouds, and precipitation) is possible through climate models. However, considering the complexity of the climate system, it is unfeasible to explain all processes occurring on Earth simultaneously. Climate models numerically solve the differential equations describing the fluid dynamics of the atmosphere and ocean on a discrete grid. Processes that are smaller than the grid-scale of the model cannot be solved directly and are thus poorly represented (IPCC ). Therefore, in order to better represent small-scale processes in the atmosphere, there has been a shift towards developing high-resolution models with smaller grid cells. Newest developments allow for very-high-resolution simulations in realistic, weather prediction mode, albeit over limited spatial domains and for short time periods only. One of the high-resolution simulations that can be used to simulate small-scale processes in the atmosphere is the ICOsahedral Non-hydrostatic Large-Eddy Model (ICON-LEM) (Dipankar et al. , Heinze et al. ). However, due to the extremely high computational cost required, this simulation-based approach can only be run for a limited amount of time within a limited area. To address this, we developed new models using emerging machine learning approaches that leverage a plethora of satellite observations providing long-term global spatial coverage up to several decades. In particular, we developed machine learning models capable of capturing the key process of precipitation formation for liquid clouds, the autoconversion process. The term autoconversion is used to describe the collision and coalescence of cloud droplets responsible for raindrop formation. This process is a key in better understanding the response of clouds to anthropogenic aerosols (Mülmenstädt et al. ). We validate the performance of our models against simulation data and then use our best model (DNN) to predict the autoconversion rates directly from satellite data.